Bayesian Case-Exclusion and Personalized Explanations for Sustainable Dairy Farming (Extended Abstract)

Eoin M. Kenny, Elodie Ruelle, Anne Geoghegan, Laurence Shalloo, Micheál O'Leary, Michael O'Donovan, Mohammed Temraz, Mark T. KeaneIJCAI 2020 · DOI

TL;DR

This paper introduces PBI-CBR, a novel case-based reasoning system for predicting grass growth in dairy farming. It tackles the challenge of user adoption in smart agriculture by combining predictive accuracy with personalized explanations, leveraging Bayesian methods for robust data cleaning and improved decision-making.

What Problem Does This Paper Address?

The agricultural sector, particularly dairy farming, faces increasing pressure to become more sustainable and mitigate its environmental impact. Artificial intelligence (AI) offers promising solutions through decision support systems (DSS), but user adoption is often hindered by a lack of trust stemming from opaque predictions. Farmers need systems that are not only accurate but also explainable, allowing them to understand why a prediction is made. This paper addresses this by enhancing an existing DSS, PastureBase Ireland (PBI), with a case-based reasoning (CBR) approach that prioritizes both predictive accuracy and personalized, case-based explanations for grass growth prediction. A key challenge in this domain is the inherent noise in farmer-entered data, which can degrade the performance of AI systems.

Methods

The core of the PBI-CBR system is its innovative use of Bayesian Case-Exclusion for maintaining the case base in a regression domain. This technique addresses the noise in farmer-provided data by using a "gold-standard" dataset (collected under controlled conditions) to inform prior beliefs about data distributions. This allows the system to identify and exclude outlier cases from the farmer-entered data, creating a cleaner "working-farm" dataset.

The paper explores several variants of this approach:

  • Control: A baseline system using all available data.
  • Exclude-2sd/3sd: Systems that exclude cases falling outside two or three standard deviations from the mean growth rates in the gold-standard dataset for a given week.
  • Transform-3sd: A system that transforms the growth rates of cases using the gold-standard distribution's parameters, effectively "normalizing" them to be more comparable to the cleaner data.

Furthermore, the paper investigates updating Bayesian priors year-on-year. This adaptive approach aims to track shifts in grass growth distributions caused by factors like climate change, using Bayesian calculations to update the model's understanding of expected growth rates over time.

For prediction and explanation, a standard k-Nearest Neighbors (k-NN) algorithm is employed. Explanations are provided by referencing similar cases from the same farm or, if unavailable, from nearby farms within the same county.

Datasets

  • Gold-Standard Dataset: A meticulously collected, noise-free dataset of grass-growth measurements spanning 28 years (1982-2010) from Teagasc's Moorepark Dairy Research Centre. This dataset is crucial for establishing baseline distributions and identifying outliers.
  • PastureBase Ireland (PBI) Dataset: This dataset comprises weekly grass-cover records from over 6,000 Irish dairy farms, collected between 2013 and 2017. This is the primary dataset used for constructing the working-farm case base after initial cleaning.

Key Insights and Ideas

  • Bayesian Priors for Data Cleaning: The paper demonstrates that Bayesian methods are not just for prediction but can be powerfully applied to clean noisy, real-world data in agricultural DSS. Using a gold-standard dataset to inform priors is a clever way to leverage existing high-quality data to improve the usability of less reliable data.
  • Trade-off Between Accuracy and Explanation: The experiments clearly illustrate the inherent trade-off between predictive accuracy (measured by Mean Absolute Error - MAE) and explanatory success (measured by %Farm-Explanation-Success - %FES and %County-Explanation-Success - %CES). Systems that retain more cases (like Transform-3sd) tend to offer better explanations but might have slightly higher errors, while aggressive exclusion can reduce errors but potentially limit explanatory options.
  • Adaptive Priors for Climate Change: The finding that updating Bayesian priors year-on-year can improve performance is critical. This suggests that AI systems in agriculture need to be dynamic and adapt to changing environmental conditions, such as those brought about by climate change, rather than relying on static models.
  • Personalized Explanations Drive Adoption: The emphasis on providing explanations by referencing similar cases from the same farm or nearby farms is key to building farmer trust. This personalized approach makes the AI's reasoning more relatable and actionable.

Key Results

  • Experiment 1 (Bayesian Case-Exclusion):
    • All exclusion and transformation methods improved prediction accuracy (reduced MAE) compared to the control system.
  • The Transform-3sd system achieved the best MAE (around 8.6 kg DM/ha/day) at a specific k value (k~35), while still offering a balanced trade-off with explanatory success at larger k values (k=1000, MAE ~8.6, %FES ~85%).
  • %FES was generally low for small k but increased significantly beyond k=50, indicating that retrieving more neighbors helps find same-farm explanations. %CES was consistently high (~100%).
  • Experiment 2 (Updating Priors):
    • Bayesian updating generally improved performance across all tested systems, though the improvement was minimal for the transform-system.
  • Explanatory success (%FES) was slightly lower in this experiment compared to Experiment 1, likely due to less training data. However, acceptable trade-offs were still achieved.

Limitations and Future Work

  • Data Size for Explanation: While Experiment 1 showed good explanatory success at k=1000, the paper notes that more data might be needed for similar success in farm-specific explanation retrieval, especially for the Transform-3sd system.
  • Less Training Data in Experiment 2: The updated priors experiment (Expt. 2) used a smaller training dataset (2013-2016 vs. 2013-2017 in Expt. 1), which likely impacted the explanatory success metrics.
  • Future Work: The authors suggest that their recent work [Temraz et al., 2020] shows these techniques can continue to deliver accurate predictions even with significant climate shifts by re-evaluating previously excluded "outlier" cases in later years. This implies a future direction of more dynamic case management and potentially incorporating a wider range of environmental factors. The potential for these techniques to improve the sustainability of dairy farming is highlighted as a key future impact.

Summary updated

BibTeX
@inproceedings{kenny2020bayesian,
  author = {Eoin M. Kenny and Elodie Ruelle and Anne Geoghegan and Laurence Shalloo and Micheál O'Leary and Michael O'Donovan and Mohammed Temraz and Mark T. Keane},
  title = {{Bayesian Case-Exclusion and Personalized Explanations for Sustainable Dairy Farming (Extended Abstract)}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.24963/ijcai.2020/657},
}
Figure